Accelerating denoising diffusion probabilistic model via truncated inverse processes for medical image segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical image segmentation requires high accuracy and uncertainty assessment for clinical decisions.
• Denoising Diffusion Probabilistic Models (DDPMs) offer potential for segmentation and uncertainty estimation but face challenges in efficiency and noise-related errors.
• Existing DDPMs in medical imaging are hindered by slow inference and prediction inaccuracies due to noise.

Purpose of the Study:

• To develop an efficient and accurate medical image segmentation method using diffusion models.
• To address the limitations of low inference efficiency and prediction errors in current DDPMs for medical image segmentation.
• To introduce a novel accelerated DDPM approach for improved segmentation accuracy and uncertainty quantification.

Main Methods:

• Proposed an Accelerated Denoising Diffusion Probabilistic Model (ADDPM) for medical image segmentation.
• Initiated the inverse process from a non-Gaussian distribution derived from pre-segmentation, terminating early with low noise.
• Integrated a separate segmentation network for pre-segmentation and a denoising network for a single-step final segmentation.

Main Results:

• ADDPM significantly reduces denoising steps (approx. 1/10th of vanilla DDPMs).
• Experimental results on four segmentation tasks show ADDPM outperforms vanilla DDPMs and other accelerated DDPM methods.
• ADDPM demonstrated superior segmentation performance and effective uncertainty estimation.

Conclusions:

• ADDPM offers a highly efficient and accurate solution for medical image segmentation.
• The method effectively reduces computational cost while maintaining or improving segmentation quality.
• ADDPM can be readily integrated with existing segmentation models to enhance their performance and provide uncertainty insights.